OpenAI 將 GPT-6.2 受保護推理與對話歷史綁定以防未授權蒸餾
OpenAI ties GPT-6.2 reasoning to chat history to block copying
OpenAI 公佈 GPT-6.2 Astra 和 GPT-6.2 Sol 的 API 防護措施,將受保護推理與產生它的對話歷史綁定,以限制未授權蒸餾。若較早訊息被改寫,API 會移除其後的受保護推理;防護亦包括對抗訓練、輸入與輸出分類器,以及限制推理跨對話和模型重用的 API 規則。
OpenAI published details on November 20, 2026 of how it fights adversarial distillation: using a frontier model's outputs without authorization to train another model that has weaker safeguards. Its main example is cyber capability that a copied model could use for attacks the original would refuse.
What changed in the API
OpenAI puts particular weight on the model's protected reasoning, the intermediate steps it generates before answering. In GPT-6.2 Astra and GPT-6.2 Sol, that reasoning is tied to the conversation history that produced it. New messages can build on earlier reasoning, but if an earlier message is rewritten, the API drops the reasoning that followed it.
Older model families with weaker safeguards are also blocked from reading the protected reasoning of Astra and Sol. OpenAI says it strengthened protections across user, workspace, organization and model-family boundaries, and closed a pathway that could have exposed another user's encrypted reasoning.
Safeguards around each request
The company describes three layers at the request level:
- Adversarial training, so the model refuses to disclose protected reasoning while still explaining things to legitimate users.
- Input classifiers that spot extraction attempts, and output classifiers that catch leaks in assistant messages and tool-call arguments, in both streaming and non-streaming responses.
- API rules limiting how protected reasoning can be reused across conversations and models.
The numbers
In adaptive black-box red-teaming, aggregated across six attacker models, success in recovering reasoning from an earlier request fell from 54.0% on GPT-6 Astra with its original safeguards to 0.2% on GPT-6.2 Astra. For the current request it fell from 22.9% to 0.9%. These are OpenAI's internal results. OpenAI says it is now applying selected improvements to GPT-6 Astra too.
Separately, an internal study found that a student model trained on exact reasoning traces generally beat one trained on reasoning summaries after supervised fine-tuning. After a light round of reinforcement learning, the exact-trace student still led at the largest inference budget tested.
Enforcement
OpenAI reports seeing several unauthorized distillation campaigns this year. Its weekly enforcement actions have covered from thousands to tens of thousands of accounts, with the rate rising sharply over the summer. It acts against whole account networks, and shares indicators of compromise with other model providers. It named DeepSeek and individuals associated with Moonshot AI as actors in earlier disclosures, and says it supports US legislation giving the government more tools against foreign extraction.
The post lists steps for other developers: check tool-call arguments for leaks, keep successful attacks as regression tests, and pair request-level controls with campaign investigations.
來源:DEV Community · dev.to